Research On Face Frontalization And Model Compression Based On Generative Adversarial Network

Many applications nowadays use frontal face images as the criteria for judging the identity of other people,but in many scenes,the person who needs to collect the face is in an uncontrolled environment and can only collect images from other angles that are not frontal.Faces,which cause the performance of many face recognition algorithms to decline in these scenarios.The use of face correction technology can improve the accuracy of person identification in these scenarios.Thanks to the rapid development of deep learning methods and the large number of easy-to-obtain annotated face images,face recognition technology in unrestricted environments has made great progress in recent years.Although face recognition technology has achieved surpassing human performance on several benchmark data sets,posture changes are still a bottleneck in many practical application scenarios.The existing methods for solving posture changes can be divided into two categories: one uses hand-designed or learned posture invariant features,and the other uses generation technology to restore frontal face images from facial images with large poses.Then use the replied facial image for face recognition.Due to the trade-off between invariance and distinguishability,the first type of method cannot effectively deal with the situation of large gestures.In the second type of methods,methods based on deep learning have achieved good results in the field of face correction in a data-driven manner.The method based on generative confrontation network has achieved good results in this field.The GAN shines in frontal face image generation,and the generated frontal face is extremely lifelike and is favored by researchers.However,its powerful image generation ability comes from the huge amount of calculation in its training and use process.The more complex the GAN structure,the calculation demand,which greatly limits its interactive deployment.In order to enhance the convenience of its deployment and reduce the computational requirements of GAN,this paper proposes a general compression algorithm,which compresses the face conversion GAN and reduces the reasoning time and model size of the generator in GAN.The experiment in this paper proves that the algorithm in this paper greatly reduces the amount of calculation compared with the original network,and the compressed GAN network still maintains a good picture quality.Finally,this article uses graphical interface programming technology,using the compressed face generation model to implement a system for face correction.The system mainly includes five functional modules,which realize the functions of generating frontal faces based on side-face photos,log query,and detection update,which facilitates people's use of the face-correction model.